FR2939264A1 - DEVICE FOR ENCODING A STREAM OF DIGITAL IMAGES AND CORRESPONDING DECODING DEVICE - Google Patents

Abstract

A decoding device of a digital image stream, comprising: a working memory (42) for already decoded image block data, a driver (40), mounted to analyze an incident flow of image data digital, and arranged to determine data to be decoded from a current block of the complete data or data-parameter type, for storing the complete data in the working memory (42) as already decoded picture block data, and for calling a decoding unit (48) with the parameter data, the decoding unit (48) responsive to the reception of parameter data by: - selecting some already decoded image blocks, forming a base of approach, which approximates the current block according to a chosen rule, - the calculation of an approximation of the current block, from a linear algebraic function of the data of the approach base, and - the storage in the working memory ( 42), for the current block, already decoded image block data derived from this approximation.

Description

INRIA 101.FRD 1 Encoding device for a digital image stream and corresponding decoding device

The invention relates to image processing and more particularly to video coding architectures.

The last ten years have seen a significant increase in the distribution of videos in electronic form. The most popular and popular format nowadays is the MPEG-2 used by DVDs. The quality of this standard has since been improved with MPEG-4 and more recently with H.264. However, as the coding architectures progressed, several problems gradually arose.

At first, it was necessary to improve the quality of the rendered images. This was achieved by gradually increasing the size of the encoded files.

This allowed to increase the encoding rate, that is to say the amount of data per second. Quite quickly, the throughput had to be limited for reasons of material limitations.

Various architectures, including those mentioned above, have therefore been developed in order to compress more accurately at a comparable rate. These architectures have resulted in a consequent increase in the cost of decoding and encoding computing time as architectures become sophisticated.

In addition, this sophistication has resulted in making the encoded information more and more critical. Indeed, the more the architectures are advanced, and the more data to be decoded represent a "digest" of information, and the more the transmission errors of these data tend to degrade the rendered image. The invention improves the situation.

For this purpose, the invention proposes a device for decoding a digital image data stream, comprising: a working memory for already decoded image block data, a driver, mounted to analyze a stream digital image data incident, and arranged to determine data to be decoded from a current block of the complete data or data-parameter type, for storing the complete data in the working memory as image block data already decoded, and for calling a decoding unit (48) with the parameter data, * the decoding unit responsive to the reception of parameter data by: - selecting some already decoded image blocks, forming a base d approach, which approximates the current block according to a chosen rule, - the calculation of an approximation of the current block, from a linear algebraic function 15 of the data of the approach base, and - the storage in the memory of working, for the current block, already decoded picture block data derived from this approximation.

The invention also relates to a device for encoding a stream of digital image data, characterized in that it comprises: a working memory for already encoded image block data, a driver, configured to analyze an incident flow of digital image data, and arranged to determine data to be encoded from a current block of the complete data or data-parameter type, to store the complete data in the working memory as data. of already encoded image blocks, and to call an encoding unit with the parameter data, * the encoding unit reacting to the reception of parameter data by: - the selection of some already encoded image blocks , forming an approach base, which approximates the current block according to a chosen rule, 30 - calculating an approximation of the current block, from a linear algebraic function of the data of the approach base, and 2

3 - storing in the working memory, for the current block, already encoded image block data and data to be decoded from this approximation.

Such devices are particularly advantageous because they offer a high image quality while allowing to implement a wide variety of encoding schemes that make compromises compression ratio / robustness extremely interesting.

Other features and advantages of the invention will appear better on reading the following description, taken from examples given by way of non-limiting illustration, taken from the drawings in which: FIG. 1 represents a schematic view of an encoder and a decoder according to the invention; FIG. 2 represents an operating flowchart of a first embodiment of an encoder of FIG. 1; FIG. 3 represents a function of FIG. 2; FIG. 4 represents an operating flow diagram of a second embodiment of an encoder of FIG. 1; FIG. 5 represents an exemplary function of FIG. 4; and Figure 6 shows an operating flow chart of an embodiment of a decoder of Figure 1.

The drawings and the description below contain, for the most part, elements of a certain character. They will thus be able not only to serve to better understand the present invention, but also to contribute to its definition, if necessary.

This description is likely to involve elements likely to be protected by copyright and / or copyright. The rights owner has no objection to the identical reproduction by anyone of this patent document or its description, as it appears in the official records. For the rest, he reserves his rights in full. FIG. 1 represents a schematic view of an encoder 2 and a decoder 4 according to the invention.

The encoder 2 receives a decoded video file 6, and transforms it into an encoded video file 85 for storage in a storage memory 10.

It will be noted first of all that, although the coding architecture and the particular elements that compose it are particularly adapted to the coding / decoding of video files, the invention is not limited to this single domain and may find its application with 10 other types of files.

Similarly, storage memory 10 means any type of storage memory, more conventional nowadays as a hard disk, or a flash memory, more complex storage systems SAN, NAS or otherwise stored through a network. Conversely, the decoder 4 receives an encoded file 8, and transforms it into a decoded video file 6 for storage in a storage memory 10.

Encoder 2 comprises a chopper 20, a buffer memory 22, a spatial approximation unit 24, a temporal approximation unit 26, and an encoding unit 28.

The decoder 4 comprises a chopper 40, a buffer memory 42, a spatial approximation unit 44, a time approximation unit 46, and a decoding unit 48.

As will become clearer later, the encoder 2 and the decoder 4 have extremely similar structures. These similarities reflect the fact that this architecture is symmetrically designed to make the decoder 4 very independent of the encoding process. Indeed, this symmetry of structure makes it possible to ensure that the decoder 4 can operate decoupled from the encoder 2, since it operates very similarly. As will be seen below, the buffers 22 and 42 may be viewed as working memories.

The invention uses a coding technique which aims to approach a block of an image by a linear combination of other known blocks.

This technique is much more effective than the "motion vectors" known from the prior art in terms of quality. In addition, as will be seen below, this technique makes it possible to reduce the amount of critical data that must undergo entropy coding. This makes the decoder 4 more independent of the transmitted encoded information and therefore more robust than in the prior art.

The spatial approximation units 24 and 44 and temporal 26 and 46 implement mathematical "matching pursuit" algorithms, which make it possible to approach an image block by a linear combination of already known blocks, while minimizing the number of frames. elements in this combination.

The matching pursuit algorithms operate by successive iterations, gradually constructing a weight vector associated with an approach base for an input vector.

This means that, to approach a given vector (for example a vector composed of the luminance and chrominance data of a given block of pixels) by a given approach base (for example set of vectors composed of luminance and chrominance data. of surrounding blocks of pixels), we look for a sequence of coefficients.

This series of coefficients is such that, by realizing a linear combination of the vectors of the base with these coefficients, we obtain an approximation of the input vector with a precision threshold chosen. The sequence of coefficients may itself be represented as a vector whose values are set iteratively.

The matching pursuit algorithms aim to produce the coefficient vector as quickly as possible, and so that it contains a maximum of zeros, which makes it possible to transmit only the non-zero values (and their index in the vector). coefficients).

Each operation of a "matching pursuit" algorithm is characterized by the index of the iteration, the approach vector calculated at this iteration, and the approximate vector corresponding to the product of the approach vector by the base of approach. It is possible to store the intermediate results of all the iterations, and the algorithm stops when the difference between the approximate vector and the given vector is less than a threshold value. By difference, we mean the measurement chosen for the algorithm, for example the Euclidean norm. There are many variations in matching pursuit algorithms, and those skilled in the art will recognize them. Other algorithms could be implemented than "matching pursuit" algorithms, which are based on a reconstruction from the surrounding block values. Those skilled in the art will also recognize them and adapt the invention if necessary.

The time approximation units use as reference blocks for the block approach base in the preceding (forward prediction) or forward prediction frames, or a mixture of both. Spatial approximation units use as reference blocks for the approach base blocks of the same image.

The notion of reference block involves cutting an image of a given file into regular blocks. This cutting is done by the cutters 20 and 40. These elements have the function of "cutting" the file into current blocks, and to process these current blocks by transmitting them to the encoding and decoding units. The cutters 20 and 40 therefore act as pilots, processing the images one by one, and block by block.

The cutting of an image being classically done from the top left corner to the lower right corner by browsing the image by line of blocks, these blocks are therefore chosen "at the top (both left and right) relative to the block and left in the block line "relative to a current block.

It follows that for a current block to be encoded / decoded by the spatial method, the blocks at the top left preceding it must be known.

This means that the leftmost and highest blocks of an image can not be encoded by this method and that they must be encoded / decoded differently for example by an "intra" encoding, others blocks of the same image, for example by applying a transformation, then a quantization followed by an entropy coding. Alternatively, these blocks can also be encoded with the time method.

It goes without saying that if the image is traveled in a different direction, it is this meaning that will define which blocks must be known for the application of the spatial method.

The Applicant has observed that, for each series of approximately 25 images, the first image is generally encoded according to the spatial method, and the following according to the temporal method.

Thus, there is only one image out of 25 that requires initialization. As described above, this initialization can be done by coding the edge data by transformation, quantization then entropy encoding, or by coding according to the time method. It follows that only the top edge and the left edge of an image over 25 in practice need to be transmitted by coding the edge data by transformation, quantization and entropy coding to achieve the initialization of the invention. . Then, the operation proceeds as described above for the rest of this image, as well as for the next 24 images.

In addition, as will be seen below, except for the blocks on the top and left edges of the image, the blocks of an image can be encoded independently by the spatial or temporal method. A marker indicating the method used will be sufficient for decoding.

As it now appears more clearly, it is therefore necessary to distinguish two types of data from the current blocks of the file 8: the complete data are not intended to be decoded and which are the data of the blocks referred to in the paragraphs previous ones; and - parameter data, which will be decoded, and which represent most of the data in file 8.

In the embodiment described here, it is the chopper 40 that makes this distinction and which: stores directly in the working memory 42 the complete data as already decoded image block data; calls the decoding unit 48 with the parameter data for their processing.

This distinction could, however, be made by the decoding unit itself. The situation is similar on the encoder side. Indeed, as we have seen above, it is known from the data to be encoded what are complete data and what are parameter data. In the embodiment described here, it is the chopper 20 that makes this distinction and which: stores the complete data directly in the working memory 22 as already encoded image block data; calls the encoding unit 28 with the parameter data for their processing.

This distinction could however be achieved by the encoding unit itself.

The encoded file 8 may further comprise as data-parameters of each block of data to be decoded and data to be decoded complementary. The encoded data and the additional encoded data will be further described with the various embodiments.

FIG. 2 represents a generic diagram of the operation of a first embodiment for the encoder 2. In this first embodiment, the architecture is based on the principle that the encoder-side and decoder-side approximation units will operate. identically.

For this, the approximation of a given block is not done on the basis of this block, but on the basis of surrounding blocks already encoded / decoded. Thus, the decoder 4 only needs to know the method that was used to calculate the approximation of a block to find it.

In this first embodiment, the encoding unit 28 calls the time approximation unit 24 in an operation 200 and the spatial approximation unit 26 in an operation 210, with the current block received from the chopper 20.

During these operations, the units 24 and 26 determine the approximations of the current block 25 according to the spatial and temporal methods. Operations 200 and 210 can be performed sequentially or in parallel.

In an operation 220, the encoding unit 28 receives these two approximations, determines the one that is closest to the current block, and stores it in the buffer memory 22. In other embodiments, the criterion of choice between these approximations can be based on a compromise between the quality of the approximation (that is to say the absolute fidelity of the approximation) and the flow (the quantity of information to be transmitted) for the complementary data (residual and / or iteration number).

The operation 220 can be followed by two optional operations 230 and 240, and the coding of the encoded data of the current block is performed in an operation 250.

In the optional operation 230, the encoding unit can furthermore choose a most favorable iteration. As we have seen, in this first embodiment, it is the neighborhood of the current block which is approached and not the current block itself.

It follows that the approximation obtained by successive iterations until reaching a fixed threshold is not necessarily the closest to the current block.

It is indeed possible that one of the iterations produced an approximation that did not approach the neighborhood with the required precision of the threshold, but approached the current block more closely than the approximation of the last iteration.

The optional operation 230 thus makes it possible to further increase the accuracy of the encoding, by choosing the one of the iterations which represents the best approximation of the current block.

Operation 230 can be performed in a variety of ways.

Thus, it can be performed sequentially with the operation 220, that is to say that in the operation 220, the most accurate method is chosen, and that in the operation 230 the most precise iteration of the method chosen in operation 220 is determined.

In another variant, the operation 230 may be incorporated in the operation 220, the iteration representing the best approximation of the current block for the two methods being selected from all the iterations of the two methods.

For carrying out operation 230, it is possible to operate after operations 200 and 210 have been executed. In this case, it suffices to store in the buffer 22 the approximation as a result of each of the iterations of each method with its index.

It then suffices for the operation 230 to access these approximations and to choose the iteration index of the best one.

In another variant, the operation 230 can be simplified by slightly modifying the operations 200 and 210 so that they store during the iterations the best approximation of the current block and its iteration index.

This can be achieved by adding in the matching pursuit algorithm a suitable buffer which stores the best approximation in process.

Then, at each end of iteration, the approximation obtained is compared to the approximation of the buffer.

If this new approximation approaches the current block better, then it is stored in the buffer. Otherwise, the approximation stored in the buffer is unchanged.

Thus, at the end of the operations 200 and 210, the buffer will already contain the best spatial approximation on the one hand and temporal on the other hand, and the operations 220 and 230 will be performed by the choice of the best of the two approximations drawn from these operations. At the end of the operation 230, the approximation of the corresponding current block is stored in the buffer memory 22. In another optional operation 240, the encoding unit 28 determines a residue and the encode. The residual is calculated by realizing the difference between the data to be encoded from the current block and the approximation of this block which has just been calculated. This encoding typically includes a DCT transform and quantization to eliminate null or weak terms. The resulting data then undergoes entropy encoding and is stored as complementary encoded data in the encoded file.

Then, the complementary encoded data is dequantized and undergoes an inverse DCT operation and is added to the current block approximation that corresponds to the encoded data to obtain encoded-decoded data.

This encoded-decoded data is then stored in the buffer memory 22 in place of the approximation of the current block. Thus, it can be considered that these data constitute new coded picture block data, which will replace the data to be encoded for the encoding of the subsequent blocks.

The purpose of this operation is to prevent quantization noise. The fact of using the encoded-decoded data for the encoding of the subsequent blocks makes it possible to avoid a drift due to the quantization noise on the residues.

Indeed, since the approximation of the current block is not transmitted to the decoder, the encoded-decoded data represents the data that will be available to the decoder 4 during the decoding.

It is therefore advantageous that these data be used to encode the following blocks.

Note that, in the case where the operation 240 is not performed, the data of the current block approximation from the operation 220 present in the memory 22 corresponds to the encoded-decoded data, or data of block of picture already encoded. Finally, in an operation 250, the type of approximation (spatial or temporal) derived from the operation 220 is encoded, and if necessary, the index of best iteration derived from the operation 230. FIG. 3 represents an example implementation of the "matching pursuit" algorithm of operations 200 and 210.

In an operation 300, the approximation unit determines an approximation neighborhood of the current block. It is this approximation neighborhood that will represent the current block in the algorithm.

In the embodiment described herein, the approximation unit comprises an approximation neighborhood selector that performs this selection from the coordinates of the current block.

The approximation neighborhood is chosen in the buffer memory 22 with blocks directly surrounding the current block. As we have seen above, this makes it possible to take into account the data available at the decoder for the encoding of the following blocks. Conveniently, if the current block has an index (x; y) in the clipping of the block image, this can be done by choosing blocks of coordinates (p; q) with: - p selected in the interval [x-kl; x], - q chosen in the interval [y-k2; y], and 25 - (p; q) different from (x; y), kl and k2 being two indices which will determine the size of the approximation neighborhood.

These indices will also affect the algorithm's calculation cost and accuracy.

In fact, in general, it can be considered that if k 1 and k 2 are too small, then the approximation neighborhood may not be representative of the current block, for example in high contrast regions.

Conversely, if k1 and k2 are too large, then the approximation neighborhood may not be more representative because it is obviously too distinct from the current block. In addition, the larger the neighborhood, the higher the computing cost will be.

Once selected, the data associated with these blocks are ordered in a vector that will form the vector to approach.

Note that this vector is called "causal neighborhood" because it is considered "causally" connected to the current block.

Next, in an operation 310, the approximation unit determines a search neighborhood of the current block. It is this research neighborhood that will serve as the basis of approach in the algorithm.

In the embodiment described here, the approximation unit comprises an approach base selector that makes this selection from the coordinates of the current block.

In the case of the spatial method, this can be done by defining a search window comprising index blocks (p; q) with: - p chosen in the interval [xm; x] for q chosen in the interval [yn; y-k2-1], and - p chosen in the interval [xm; xk l - l] for q chosen in the interval [y-k2; y], m and n being indices that define the size of the search window. This search window takes into account the fact that, for decoding with the spatial method, only the blocks of the current image that have already been decoded will be accessible to the decoder.

Therefore, to ensure symmetry of the architecture, encoding side, only blocks "already encoded" must be used. These "already encoded" blocks are those that are above the causal neighborhood, and "left" in the same line, as described above.

In the case of the temporal method, the selection of the search window can be done by choosing blocks in a rectangle of sides m and n, centered on the current block, taking into account the edges of the image, in one or several previous images (or following depending on the method) that have already been encoded. In the same way as for the spatial method, the indices m and n will affect the calculation cost of the algorithm and its accuracy.

Once the search window has been defined, the blocks are arranged in a search neighborhood in the form of a matrix whose: - each column comprises the samples of kl * k2 blocks, that is to say the same number of blocks than the added causal neighborhood of the block to be encoded, and each column comprises blocks of the search window, chosen for example by a sliding window starting from the coordinate block [xm; y-n] and which traverses the search window from top to bottom and from left to right as defined above. The sliding window has a size kl * k2.

Thus, the matrix comprises the data of the already encoded blocks selected as the search neighborhood for the block to be encoded, and in each column the last elements correspond to the data of the block to be encoded.

Finally, the matrix is reduced to a number of lines corresponding to that of the causal neighborhood. This is in fact to remove the elements of the sliding window that correspond to the data of the block to be encoded. These removed items are stored for later.

Then, in an operation 320, the algorithm of "matching pursuit" as such is performed as described above by a calculator, with as input vector the causal neighborhood, and as the approach base the search neighborhood. This operation can be accompanied, as described above, with storing the best approximation of the current block for the optional operation 230.

Finally, in an operation 330, the matrix corresponding to the search neighborhood from the search window is restored with the values that have been stored at 310, and this matrix is multiplied by the approximation (weight vector) obtained at 320. .

Then, the last terms corresponding to the block to be encoded of the resulting vector are retrieved and returned as an approximation of the current block.

FIG. 4 shows a generic diagram of the operation of a second embodiment for the encoder 2.

In this second embodiment, the architecture is based on the principle that the encoder-side and decoder-side approximation units will operate in a different manner, but with data inducing an identical result.

For this, the approximation of a given block is performed for the given block and more for blocks in a causal neighborhood thereof. Here, the decoder 4 will need to know the method that was used to calculate the approximation of a block to find it, as well as the weight vector that makes it possible to approach the given block.

In this second embodiment, the operations 400, 410 and 420 are essentially identical to the operations 200, 210 and 220, except that the approximation is slightly different as will appear better with FIG. 5. FIG. implementation of the "matching pursuit" algorithm of operations 400 and 410. As mentioned above, this implementation is slightly different, as it is no longer based on an approximation of the surrounding blocks, but on an approximation of the block itself. Thus, an operation equivalent to the operation 300 is not necessary, and the approximation begins with an operation 510 of determining the search neighborhood.

It is this research neighborhood that will serve as the basis of approach in the algorithm. In the case of the spatial method, this can be done by defining a search window comprising index blocks (p; q) with: - p chosen in the interval [xm; x] for q selected in the interval [0; y-1], and - p chosen in the interval [xm; x-1] for q = y, 10 m and n being two indices which will determine the size of the search neighborhood and therefore the calculation cost of the algorithm and its accuracy.

Indeed, in general, it can be considered that if m and n are too small, then the search neighborhood may not be representative of the current block, for example in regions of high contrast. Conversely, if m and n are too large, then the search neighborhood may not be more representative because it is obviously too distinct from the current block. In addition, the larger the neighborhood, the higher the calculation cost will be.

In the case of the temporal method, this can be done by selecting blocks in a rectangle of sides m and n, centered on the current block, taking into account the edges of the image, in the preceding (or subsequent) images. method) that have already been encoded.

Once the search window has been defined, the blocks are arranged in a search neighborhood in the form of a matrix whose each column contains the values of a block. Operation 510 is therefore functionally close to operation 310, except that it is simpler since it is not necessary to substitute the causal neighborhood for the current block.

This simplifies the establishment of the search neighborhood and makes it possible to include in addition blocks corresponding to the causal neighborhood in the case of the spatial method. 25 15 20 25 30

Then, in an operation 520, the algorithm of "matching pursuit" in itself is realized, with as input vector the current block, and as the approach base the search neighborhood.

As seen here, the operation 520 is close to the operation 320, except that, as mentioned above, it is the current block itself that is approached by the search neighborhood and plus the causal neighborhood of the current block.

Finally, in an operation 530, the approximation obtained is returned, that is to say the approximated vector 10 with the corresponding vector of coefficients.

It thus appears that the methods of approximation of the first and second embodiments are very close, and that their difference relates mainly to the element that is approached. As mentioned above, the operations 400 and 410 are followed by the operation 420. This operation is identical to the operation 220 of FIG. 3 and aims at choosing the one of the two approximations (spatial or temporal) which is the best in function established criteria (block loyalty, quality / debit compromise, etc.).

Then, an optional operation 440 can be performed. This operation is similar to the operation of encoding the residue 240 of FIG. 2. But this operation must take into account that it is the weight vector that is transmitted in this embodiment, and not only the encoding method. .

It is a question here of operating in the same way, namely to encode the residue by transformation, then quantization and entropy coding for storage as complementary encoded data. Then, decode entropically, then dequantize, and do the inverse transformation. In parallel, it is planned to quantify and dequantize the approximation derived from operation 420 to take into account the quantization noise on this element. Finally, add these two elements, and store it in the buffer memory 22 to replace the approximation taken from the operation 420 and thus improve the encoding of the following blocks. Finally, in an operation 450, the type of approximation (spatial or temporal) derived from the operation 420 and the corresponding approximation vector containing the weights of the linear combination of the search base are entropically coded and stored as data. encoded picture block. In the case where the operation 440 has not been performed, the operation 450 provides for quantifying and dequantizing the approximation (weight vector) derived from the operation 420 to take account of the quantization noise on this element, and store it in buffer 22 to replace the approximation from operation 420 and improve the encoding of subsequent blocks.

Thus, as for the first embodiment, the following are obtained: data to be decoded, which is the type of approximation (spatial or temporal) derived from the operation 420 and the corresponding approximation vector containing the weights of the combination linear of the search base that are encoded by quantization and entropy coding; and - already encoded image block data, which are the encoded-decoded data (taking into account the transmission of the residuals or not) and which are stored in the buffer memory. As seen here, the main difference with the first embodiment is that it is the block itself which is approximated, and that this approximation is transmitted as encoded data with a marker of the method used to select the blocks. which were used to calculate this approximation. It is sufficient for the decoder to multiply the vector containing the weights of the linear combination

20 by the search base designated by the method marker to achieve the linear combination with the already decoded data of the blocks of the search neighborhood to find the approximation of the original block.

Note that in the two embodiments described, a flag or other data may be recorded in the file 8 to indicate which embodiment was used for encoding, as well as other information, such as the threshold convergence of the pursuit matching algorithm, the dimensions of the images etc ...

FIG. 6 represents an operation performed by the decoder 4 to decode the file 8.

In a first operation 600, the decoder 4 calls the chopper 40 to cut the images into current blocks and process them. In this same operation, the chopper 40 can read a flag or other data file 8 which indicates what is the embodiment that was used to perform the encoding.

In both embodiments, this can also be derived from the approximation method marker. The chopper 40 then sends the data to be decoded in the decoding unit 48.

The decoding unit 48 may then send the data into either of the spatial approximation unit 44 and the time approximation unit 46, with a flag indicating how the linear combination of the data to be decoded must be made with the search neighborhood.

Then, in an operation 610, the unit 44 or 46 that received the data to be decoded operates to determine an approximation of the current block.

In the first embodiment, the unit 44 or 46 operates according to FIG. 3 as the unit 24 or 26, using the already decoded data present in the buffer 42.

In the second embodiment, as described above, the unit 44 or 46 realizes a linear combination of the already decoded data of the blocks of the search neighborhood in the buffer 42, to find the approximation of the original block, with a vector weight drawn from the data to be decoded.

This linear combination is in fact a direct calculation of the block to be decoded, without a matching pursuit algorithm.

The resulting approximation is transmitted to the decoding unit 48.

Then, in an optional operation 620, the decoding unit 48 decodes the additional decoded data including the residue when it has been transmitted, and adds them to the approximation received from the operation 610 in an operation 630 to recover. the decoded block.

Finally, in an operation 640, the decoded block is recorded both in the decoded file 6 thus reconstituted, and in the buffer memory 42 for the decoding of the subsequent blocks.

Some of the elements presented above can be implemented in various ways: the chopper can have a role of clipping the images into blocks only, while another element plays the role of a driver which processes the image flow image by picture ; the approximation neighborhood selector (or causal neighborhood), and the search neighborhood selector can be made in a single element which receives appropriate parameters, such as the encoding / decoding mode and the corresponding selection parameters (x, y, k 1, k2, m and n); - The selectors described here always select blocks at the top and left of the current block, so that the conditions at the edges of the image are limited. Other types of selection, especially at the top and to the right of the current block could be made taking into account these boundary conditions; The calculator of the spatial and temporal approximation units can be shared in each embodiment by these units since it performs the same function; 10 15

The spatial and temporal approximation units can therefore be separate or be a single computing unit that receives the necessary parameters; other methods than transformation coding, quantization followed by entropy coding may be used where described; - Quantization / dequantization can be avoided by sacrificing a bit of performance in terms of flow; the indices of the selectors could be varied to calculate other approximations, and the parameters of the best-performing approximation could be encoded with the encoded data; the other parameters such as the convergence threshold of the matching pursuit algorithm or the size of the images or others could be implicit for the decoder, or else taken from the file (for example from the weight of the file or its bitrate); the two embodiments could be combined for blocks of the same file 6; the selection of the blocks in the search neighborhood can be guided by a classification of these blocks according to the characteristics of the texture they contain according to the uniform area, the textured area, the predominantly horizontal outline, the vertically dominant outline; only the blocks that belong to the same class as the block to be predicted will be selected, with significant gains on the time of calculation of the approximation; the linear combination can be realized by carrying out a post-treatment such as a smoothing of the data of the search neighborhood, for example by applying a filter (for example a Gaussian kernel), or by applying geometric transformations to the data of the search neighborhood (relying on the detection of contours for example).

Several embodiments have been described above, with optional features. Whenever possible, the combination of these features with each of these embodiments has been described.

However, it is the combination of the features of the described embodiments that is within the scope of the invention.

It must therefore be understood that all the combinations of the embodiments between them and optional features are described in the present description, as well as their equivalents, even when this has not been explicitly described for the sake of simplicity.

The invention also relates to a method for decoding a digital image data stream, comprising: a. analyzing an incident flow of digital image data, for determining data to be decoded from a current block of type complete data or data-parameters, b. storing the complete data in a working memory (42) as already decoded image block data, c. decode the parameter data, step c. comprising: i. select some already decoded image blocks, forming an approach base, which is close to the current block according to a chosen rule, ii. calculate an approximation of the current block, from a linear algebraic function of the data of the approach database, and iii. storing in the working memory (42), for the current block, already decoded picture block data derived from this approximation.

The decoding method may furthermore have the following characteristics: step c.i. comprises selecting an approximation neighborhood, including already decoded picture block data corresponding to an immediate neighborhood of the current block, and step c.ii. comprises selecting already decoded image block data corresponding to an expanded neighborhood of the current block, the calculation comprising executing said linear algebraic function of the approach database data to approximate the approximation neighborhood data; * step c.ii. comprises performing said linear algebraic function a number of times drawn from iteration data that comprise the parameter data; * step c.i. comprises selecting an approximation neighborhood, comprising already decoded picture block data corresponding to an immediate neighborhood of the current block, step c.ii. comprising performing said linear algebraic function of the data of the approach database to approximate the data of the current block; * execute step c. in a decoding mode selected from the group consisting of spatial decoding and time decoding derived from decoding mode data which comprise the parameter data; * step c. comprises selecting already decoded image block data in the current image in the spatial mode, and already decoded image block data in images contiguous to the current image in the time mode; * step c.iii. comprises combining the current block approximation with corresponding complementary decode data to compute the decoded picture block data of the current block.

The invention also relates to a method of encoding a digital image data stream, comprising: a. analyzing an incident flow of digital image data, to determine b's. data to be encoded from a current block of the complete data or data-parameter type, storing the complete data in a working memory (42) as already decoded picture block data, c. encode the parameter data, step c. comprising: i. select some already encoded image blocks, forming a base of approach, which approximates the current block according to a chosen rule, ii. computing an approximation of the current block, from a linear algebraic function of the approach base data, and iii. storing in the working memory (42), for the current block, already encoded picture block data and data to be decoded from this approximation. The encoding method may further have the following characteristics: step c.i. includes selecting an approximation neighborhood, including already encoded image block data corresponding to an immediate neighborhood of the block

Current, and step c.ii. comprises selecting already encoded image block data corresponding to a widened neighborhood of the current block, and executing said linear algebraic function with the data of the approach database to approximate the approximation neighborhood data; * step c.iii. includes determining iteration data and recording it in the data to be decoded; * step c.i. comprises selecting an approximation neighborhood, comprising already encoded picture block data corresponding to an immediate neighborhood of the current block, and step c.ii. comprises performing said linear algebraic function with the data of the approach database to approximate the data of the current block; * the stages c.i. and c.ii. are executed for each current block according to a spatial mode and a time mode, and the step c.iii. comprises choosing an approximation from the resulting approximations, and encoding in the data to be decoded decoding mode data indicating the mode of the chosen approximation; * step c. comprises selecting already decoded image block data in the current image in the spatial mode, and already decoded image block data in images contiguous to the current image in the time mode; * step c.iii. comprises, for at least some of the current blocks, calculating additional data to be decoded from the data to be decoded and corresponding already encoded picture block data from these blocks, and storing them with the corresponding data to be decoded.

Claims (14)

Revendications1. Device for decoding a digital image data stream, comprising: * a working memory (42) for already decoded picture block data; * a driver (40) mounted for analyzing a data incident stream of digital images, and arranged to determine data to be decoded from a current block of the complete data or data-parameter type, for storing the complete data in the working memory (42) as image block data already decoded, and for calling a decoding unit (48) with the parameter data, * the decoding unit (48) responsive to receiving parameter data by: - selecting some already decoded picture blocks, forming an approach base, which approximates the current block according to a chosen rule, - calculating an approximation of the current block, from a linear algebraic function of the data of the approach base, and - the storage in the working memory (42), p For the current block, already decoded image block data derived from this approximation. 2. Device according to claim 1, characterized in that the decoding unit comprises: an approximation neighborhood selector, arranged to select already decoded picture block data corresponding to an immediate neighborhood of the current block, and a basic approach selector, arranged to select already decoded image block data corresponding to a widened neighborhood of the current block, and a calculator, arranged to execute said linear algebraic function with the database data. approach to approaching neighborhood approximation data. 25 3. Device according to claim 2, characterized in that the data-parameters comprise iteration data, and in that the computer is arranged to perform said linear algebraic function a number of times drawn iteration data. 4. Device according to claim 1, characterized in that the computing unit 30 comprises: an approach base selector, arranged to select already decoded image block data corresponding to an enlarged neighborhood of the current block, and a calculator arranged to execute said linear algebraic function of the data of the approach database to approximate the data of the current block. 5. Device according to one of claims 2 to 4, characterized in that the parameter data comprises decoding mode data indicating a decoding mode selected from the group comprising a spatial decode and a time decode. Device according to claim 5, characterized in that the approach basic selector is arranged to select image block data already decoded in the current image in the spatial mode, and image block data. already decoded in images contiguous to the current image in the time mode. 7. Device according to one of the preceding claims, characterized in that at least some of the parameter data further comprise complementary data to be decoded, and in that the decoding unit is arranged to combine the approximation of the current block. with corresponding complementary decoder data for calculating the decoded picture block data of the current block. 8. Device for encoding a digital image data stream, characterized in that it comprises: a working memory (42) for already encoded image block data, a driver (40) , mounted to analyze an incident flow of digital image data, and arranged to determine data to be encoded from a current block of the complete data or data-parameter type, for storing the complete data in the working memory (42) in as already encoded picture block data, and to call an encoding unit (48) with the parameter data, * the encoding unit (28) responsive to the reception of parameter data by: - the selection of certain already encoded image blocks, forming an approach basis, which approximates the current block according to a chosen rule, - calculating an approximation of the current block, from a linear algebraic function 30 of the data of the base of approach, and- the storage in the memory of the work (42), for the current block, already encoded image block data and data to be decoded from the approximation of the current block. 9. Device according to claim 8, characterized in that the encoding unit comprises: an approximation neighborhood selector, arranged to select already encoded image block data corresponding to an immediate neighborhood of the current block, and a basic approach selector, arranged to select already encoded image block data corresponding to a widened neighborhood of the current block, and a calculator, arranged to execute said linear algebraic function with the database data. approach to approaching neighborhood approximation data. 10. Device according to claim 9, characterized in that the encoding unit is arranged to pull the calculator iteration data, and to encode said iteration data in the data to be decoded. 11. Device according to claim 8, characterized in that the computing unit comprises: a basic approach selector, arranged to select already decoded image block data corresponding to a widened neighborhood of the current block, and a calculator, arranged to execute said linear algebraic function of the data of the approach database to approach the data of the current block. 12. Device according to one of claims 9 to 11, characterized in that the encoding unit is arranged to call the computer by selecting the image block data already encoded according to a spatial criterion and according to a time criterion, for selecting the approximation of the current block among the resulting approximations, and for encoding in the data to be decoded encoding mode data indicating the criterion used to generate said approximation of the given block. Apparatus according to claim 12, characterized in that the encoding mode data indicates that the approach base selector has selected image block data already encoded in the current image for the spatial criterion, and image block data already encoded in images contiguous to the current image for the temporal criterion. 14. Device according to one of claims 8 to 13, characterized in that, for at least some of the current blocks, the encoding unit is arranged to calculate additional data to be decoded from the data to be decoded and data. corresponding already encoded image blocks of these blocks, and to store them with this data to be decoded.